Revised · 12 August 2026: the original article was written when Runtime was the clearest concrete product. Subsequent Host, Harness, World, Computing, Studio, Finance, Game, and Security experiments changed the system enough that the original explanation was no longer a good public entry point. This revision keeps the original question and replaces the outdated system description.
A model can propose. A machine can execute. An external service can change. A domain can decide what the change means. Research can generalize a pattern. A website can explain it. Those are different jobs.
Start with one interrupted task
An Agent changes policy.py and starts the tests. The browser connection disappears before the result comes back. Five minutes later, another Agent takes over.
What should the new Agent ask?
- What work had actually been accepted?
- Did the test process run?
- If something external was called, did that external action commit?
- Does the old evidence still apply to the repository and environment that exist now?
- Is the larger task complete, or did only one execution finish?
A transcript may contain guesses about all five. It is not automatically authoritative for any of them.
The basic Ordivon problem is not that Agents forget words. It is that different facts survive and fail under different authorities.
The smallest useful picture
The current system can be understood without learning the object model first:
domain / participant
decides what the work means
↓
Host
keeps durable responsibility across attempts
↓
Harness
runs one bounded episode of intelligence and Tools
↓
Runtime
proves what happened on the local machine
↓
external owners / World relations
establish what happened outside the machine
↓
domain / participant
re-evaluates the current consequenceComputing stands beside that trajectory and asks which distinctions survive across projects. Studio and Web turn the work into human-facing artifacts and test what the rendered encounter actually communicates. Finance, Game, Security, and Human own their own domain meanings.
No application is required to use every box. The point is the division of responsibility, not one universal deployment diagram.
Why the model is not the owner
A model is excellent at reconstructing plausible explanations. That is exactly why it should not silently become the authority for facts that already have a stronger owner.
If Runtime has a recorded Job, the model should not infer from prose whether the process ran. If an exchange or cloud provider has a receipt, the model should not replace it with confidence. If the repository changed after an old experiment, the old experiment can remain historically valid while no longer governing the current decision.
Ordivon therefore treats intelligence as replaceable. A stronger model can arrive tomorrow without requiring the system to rewrite yesterday's physical evidence or accepted responsibility.
The boundaries are easier to see as non-implications
Much of Ordivon's recent research can be compressed into statements about what one true fact does not prove:
These separations prevent a common class of Agent-system error: taking a fact that is genuinely known and extending it one step further than its evidence allows.
What each core project is for
Host: preserve responsibility
If the first Agent disappears, someone still needs to know what objective was accepted, what remains unresolved, and where the next continuation begins. Host keeps the smallest durable semantic state whose loss would change current or recoverable responsibility.
Harness: run one episode of intelligence
The Harness decides what one Agent Run sees, which Tools it can address, how Provider and Tool continuations are represented, and when the Run stops. It should not become the permanent Task database merely because the model interacted through it.
Runtime: preserve physical execution truth
Runtime owns exact Workspaces, Jobs, Attempts, immutable inputs, process state, Artifacts, cancellation, restart recovery, and reconciliation. A successful Runtime Job is strong evidence about the machine. It is deliberately weaker evidence about the meaning of the larger task.
World: preserve the relationship neither endpoint can reconstruct alone
World was once much broader. High-pressure deletion removed research-only routing and path abstractions that current owners did not need. What survives is narrow: external observation availability, provider/consequence binding, unresolved-consequence reconciliation, and other cross-owner relations whose loss would make continuation ambiguous.
The domain still decides meaning
Infrastructure can tell Finance that one order request was admitted exactly once. It cannot decide whether the trade was a good capital decision. Runtime can prove that a Game simulation ran. It cannot decide what counts as strategic agency. Harness can preserve a Security Agent's unresolved unknown. It cannot decide what the adversarial evidence means.
This is why Ordivon is a family rather than one giant schema. Finance, Game, Security, Human, and other domains retain meanings that cannot be inferred from generic execution machinery.
Computing tests which lessons deserve to travel
Cross-project similarity is tempting. Two projects both have an UNKNOWN-like state; perhaps we should build a universal uncertainty service. Three projects persist state; perhaps they need one persistence framework.
Computing's job is to resist that jump.
A shared law is attacked through deletion, ablation, strong baselines, counterexamples, and independent domains. Even when a semantic distinction repeats, the implementation may remain owner-local. Recent EX3–EX7 work went one step further and tested how those laws should be explained: richer causal cards and relation grammars did not improve the bounded Agent decisions over compact causal prose.
Studio and Web test the explanation as an artifact
Correct source facts do not guarantee a correct public encounter.
Studio R4–R6 showed three different failures: richer perception can add features without adding explanatory power; a plausible semantic statement can fail because it cannot point back to the evidence; and a searched creative winner can look statistically extraordinary under pure noise.
Web therefore owns presentation and encounter evidence, not technical truth. It asks a different question: given the facts supplied by their owners, what does the actual rendered page make visible, imply, or hide?
What Ordivon is not
Ordivon is not an attempt to put every Agent behind one workflow engine, one ontology, one world database, or one policy layer. Several of its strongest results are deletions or decisions to add nothing.
The project family earns a shared mechanism only when a real failure survives stronger ordinary alternatives and materially different workloads.
The current thesis
The current Ordivon thesis is narrower than “build an autonomous computer” and stronger than “save Agent memory”:
Give each consequential fact the smallest owner that can prove it, keep the Agent able to address the distinctions needed for its decision, and require fresh admission before evidence becomes current consequence.
Everything else remains negotiable.
Work should be able to outlive the intelligence that started it without outliving the evidence that makes it true.
Continue
Three routes into the evidence
- From Tokens to Work — follow one Agent action through the responsibility boundaries.
- The Shorter Explanation Won — see how the explanation itself was experimentally tested.
- World Got Smaller—and Got Clearer — see deletion used as an architectural falsifier.